Remote work: what the trials actually show.

Few workplace questions draw more confident opinion per unit of evidence than remote work. Executives cite culture; employees cite commutes; both sides wave surveys built to find what their sponsors hoped. But under the noise, a real experimental literature has been building for a decade — actual randomized trials, run inside real companies, with output measured rather than self-reported. Read together, the trials tell a story more precise and more interesting than either camp’s talking points. The effect of remote work is not a number. It is a function — of the job, the design, and who selects into it.

This article walks through the anchor experiments in order. First, the call-center trial that made remote work respectable to academics. Then the hybrid randomized trial that changed the default for knowledge work. Then the studies that found real costs, and the network data that explains when and why the costs arrive. The through-line is worth stating up front: every result makes sense once you ask what the office was actually providing. For learning organizations, the most important answer turns out to be one nobody was managing.

Before the pandemic, there was a call center

It matters that the first serious experiment happened in a call center. Call centers are the fruit fly of personnel economics: output is counted automatically, quality is monitored, tasks are individual, and nothing about the job depends on hallway luck. If remote work was ever going to look good under measurement, it would be here. That makes the size of the gains — and the places where later studies found losses — easier to read.

The founding experiment predates the remote-work wars by half a decade. Nicholas Bloom and colleagues persuaded Ctrip, a 16,000-employee Chinese travel agency, to run a true lottery. It randomized 249 call-center volunteers between working from home four days a week and working in the office. The trial ran nine months, with identical pay, equipment, and performance metrics (Bloom, Liang, Roberts & Ying, 2015).

The result made the paper famous: home workers’ performance rose about 13%. The breakdown matters more than the headline. Roughly nine percentage points came from working more minutes per shift — fewer breaks, fewer sick days, no commute bleeding into the workday. About four points came from handling more calls per minute, which the authors credit to a quieter setting. Attrition — people quitting — fell by half, and reported work satisfaction rose. And one result was quieter and darker: at the same performance level, home workers were promoted less often (Bloom, Liang, Roberts & Ying, 2015).

The experiment carried a second lesson that took years to sink in. When the trial ended, Ctrip let everyone choose. Half the home workers chose to return to the office; half the office workers wanted to switch home. After this re-sorting, the performance advantage of home work roughly doubled. Selection, in other words, is half the story. The gains from remote work depend hugely on who chooses it — a fact that haunts every non-experimental comparison in this literature.

The hybrid trial that changed the default

The pandemic made the question universal, but the cleanest answer landed only in 2024. Bloom, Han and Liang randomized 1,612 engineers, marketing and finance employees at Trip.com. One group worked five days in the office; the other worked a hybrid week of three days in, two days home, for six months (Bloom, Han & Liang, 2024). These were university-educated knowledge workers — the group the argument is actually about.

On every performance measure the company tracked — reviews, promotions over the next two years, and, for engineers, code output — the hybrid group was statistically indistinguishable from the office group. What moved was retention. Quit rates fell by roughly a third, with the largest drops among women, employees with long commutes, and non-managers. Job satisfaction rose. Managers had predicted before the trial that hybrid would hurt productivity; by the end, they had revised their estimates to roughly zero (Bloom, Han & Liang, 2024). For a firm, a free benefit of this size is rare: replacing a knowledge worker typically costs a large slice of a year’s salary, and the trial cut quits by a third at no cost.

Be precise about what this trial licenses. It tested hybrid — structured, majority-office, coordinated days — not full remote. It ran in a firm with strong performance measurement. And its subjects were experienced professionals, not new hires. Within those bounds, it is about as clean as organizational evidence gets. The setup employees prefer cost the employer nothing measurable — and saved it turnover.

Where full remote shows its costs

If the story ended with Ctrip and the hybrid trial, remote work would look like a free lunch with a scheduling detail. The rest of the literature is where the bill arrives. It arrives exactly where work stops looking like a call queue: in jobs where the unit of production is a conversation, a handoff, or a junior watching a senior think out loud.

Gibbs, Mengel and Siemroth studied more than 10,000 IT professionals at a large Asian services firm as it moved to full remote during 2020. They had rich personnel data on output and time use. Output fell modestly while hours rose sharply — meetings multiplied, focus blocks broke up, coordination ate the day. Measured productivity fell by 8 to 19%, depending on the specification (Gibbs, Mengel & Siemroth, 2023). Hit hardest were employees with children at home and, tellingly, newer employees — the people who depended most on informal help. This was not a call center: the work was interdependent, and the interdependence is what got more expensive.

Emanuel and Harrington added the selection half of the ledger. Studying call-center workers at a Fortune 500 retailer, they found remote workers really were less productive than on-site peers. But a large share of the gap was who took the remote jobs, not the remote setup itself: the workers who chose remote roles were, on average, the less productive ones. The study also documented a cost that individual output metrics never capture. When senior workers went remote, the feedback and informal training received by junior colleagues dropped (Emanuel & Harrington, 2024). The people who pay for a senior engineer’s quiet home office can be the juniors who no longer overhear them.

The upside case is just as specific. Choudhury, Foroughi and Larson studied U.S. patent examiners — highly independent professionals whose output can be counted objectively — when they gained the right to work from anywhere. Output rose about 4.4%, with no measurable drop in quality, as examiners moved to cheaper places they liked (Choudhury, Foroughi & Larson, 2021). Geographic freedom works beautifully when the work travels whole.

What the network data shows

Individual output studies, however clean, treat the company as a collection of soloists. The next piece of evidence watched the orchestra. It is the study that should most cool enthusiasm for permanent, uncoordinated remote work, exactly because it measured what the trials cannot.

That study, the most important non-experimental evidence in this literature, explains the mechanism behind the collaborative costs. Yang and colleagues analyzed the communication metadata of 61,182 Microsoft employees before and after the firm’s 2020 shift to remote work. The collaboration network became more static and more siloed. Ties across groups and bridging connections fell, new ties formed less often, and communication shifted from live, information-rich channels toward text answered on a delay (Yang et al., 2022).

None of this shows up in this quarter’s output. Weak ties and cross-team bridges are how companies move knowledge, spot chances, and fold in newcomers — slow assets that lose value quietly. The network study is the connective tissue of this literature. It explains why individual-work trials find gains while interdependent settings find losses. And it names the specific channel — spontaneous connection across boundaries — that a distributed company must rebuild on purpose or lose. Earlier evidence points the same way from inside a single room: even among police radio operators, sitting near teammates measurably sped coordination on urgent tasks (Battiston, Blanes i Vidal & Kirchmaier, 2021).

The commute dividend, and why the arrangement stuck

One more body of evidence explains why this argument will not be settled by decree, whatever any single firm announces. Barrero, Bloom and Davis surveyed tens of thousands of American workers, repeatedly, through and after the pandemic. They documented both the scale of the stakes and the direction of travel. The average remote day saves roughly seventy minutes of commuting and grooming time, and a meaningful share of that flows back into work. The stigma around working from home collapsed during the pandemic. And firms’ own investments — in equipment, process, and norms — made the setup more productive than its improvised March-2020 version (Barrero, Bloom & Davis, 2021).

Their conclusion sits in their title: working from home will stick. For anyone designing companies, the practical upshot is that the question stopped being whether to allow flexible work. The labor market has priced it in, and the hybrid trial shows that refusing it simply donates retention to competitors (Bloom, Han & Liang, 2024). The question became how to run the setup so its known costs — which concentrate in coordination and junior development — are engineered against, not absorbed silently.

The selection problem in every comparison

Before drawing conclusions, one reading instruction. It separates the evidence in this article from nearly everything else published on the topic.

The instruction: mind selection. That is the method thread running through all of this, and it disqualifies most of what gets cited in the argument. People who choose remote work differ from people who don’t — in productivity, in family situation, in how much they value the setup. Mas and Pallais ran a field experiment inside real hiring. The average applicant would give up about 8% of wages for the option to work from home — but the average hides an enormous spread. Many workers value it at nearly nothing, while a sizable minority value it enormously (Mas & Pallais, 2017).

That spread is why “remote workers in our company perform worse” and “remote workers in our company perform better” are both routinely true — and both nearly meaningless. Ctrip’s re-sorting doubled the effect; Emanuel and Harrington’s remote hires were less productive as a group before they took a single call. Any policy conclusion drawn from comparing volunteers to non-volunteers inherits the selection, not the treatment. The trials exist precisely because the observational numbers cannot be trusted. And the survey evidence on preferences says the stakes are big enough that firms will keep offering the setup regardless — which makes designing it well the only real question (Barrero, Bloom & Davis, 2021).

What the evidence doesn’t show

The experimental record is young and bounded, and five limits matter for anyone setting policy from it:

• No long-run career RCTs. The promotion penalty in the Ctrip trial and the mentorship losses in the call-center data are warning lights, but nobody has run the ten-year experiment on careers, networks, and skill growth. The costs that worry economists most are exactly the slow ones.
• Innovation outcomes are unresolved. The network data shows the channels thinning (Yang et al., 2022); it does not measure whether fewer bridges produced fewer good ideas. That link remains inference, not evidence.
• The trials cluster in measurable jobs. Call handling, code, patents — work with countable output. Collaborative creative work, management itself, and roles where output resists measurement are underrepresented, and they are where the skeptics’ concerns concentrate.
• Individual output is not firm productivity. A quieter worker handling more calls can coexist with an organization that integrates newcomers worse and moves knowledge slower. Few studies measure the firm-level aggregate.
• Junior development is the weakest flank. The evidence that proximity feeds early-career learning — dropped feedback when seniors go remote (Emanuel & Harrington, 2024), newer employees hit hardest by coordination costs (Gibbs, Mengel & Siemroth, 2023) — is consistent and worrying, and no trial yet shows how to fully replace it.

Where the evidence stops

1. 1No long-run career RCTs
2. 2Innovation outcomes are unresolved
3. 3The trials cluster in measurable jobs
4. 4Individual output is not firm productivity
5. 5Junior development is the weakest flank

© 2026 FUTURE PROOF™

What this means for practice

Read as a body, the trials support a policy neither camp campaigned for — usually the mark of evidence doing its job. Hybrid, organized around coordinated in-office days, is the evidence-backed default for knowledge work: performance-neutral, retention-positive, preferred by employees (Bloom, Han & Liang, 2024). Full remote is a specialist tool. It is excellent for independent, measurable work and for reaching talent that cannot relocate (Choudhury, Foroughi & Larson, 2021); it is expensive for interdependent work and junior development. And whatever the setup, the network evidence hands management a new job. The spontaneous infrastructure of learning — overheard corrections, shoulder-tap questions, watching a senior work — no longer happens by default: it either becomes deliberate and measurable, or it stops happening.

Concretely, the playbook the evidence supports has four moves. Anchor hybrid on coordinated office days — the benefit of presence is other people being present, so uncoordinated flexibility buys the costs of both worlds. Measure output, not visible effort: every trial in this literature depended on exactly that, and the Ctrip promotion penalty shows what happens when firms measure presence instead (Bloom, Liang, Roberts & Ying, 2015). Protect junior-senior contact on purpose — scheduled reviews, paired work, explicit teaching time — since that is the channel the data shows failing first (Emanuel & Harrington, 2024). And build the knowledge base that makes questions answerable without a tap on the shoulder.

That last point is where remote-work policy quietly becomes learning policy. An office was, among other things, an ambient training system. Spread the workforce out and the ambient system switches off. That is survivable — but only for companies that replace it with something explicit.